A general structure-property relationship to predict the enthalpy of vaporisation at ambient temperatures

Abstract

The vapour pressure is the most important property of an anthropogenic organic compound in determining its partitioning between the atmosphere and the other environmental media. The enthalpy of vaporisation quantifies the temperature dependence of the vapour pressure and its value around 298 K is needed for environmental modelling. The enthalpy of vaporisation can be determined by different experimental methods, but estimation methods are needed to extend the current database and several approaches are available from the literature. However, these methods have limitations, such as a need for other experimental results as input data, a limited applicability domain, a lack of domain definition, and a lack of predictive validation. Here we have attempted to develop a quantitative structure-property relationship (QSPR) that has general applicability and is thoroughly validated. Enthalpies of vaporisation at 298 K were collected from the literature for 1835 pure compounds. The three-dimensional (3D) structures were optimised and each compound was described by a set of computationally derived descriptors. The compounds were randomly assigned into a calibration set and a prediction set. Partial least squares regression (PLSR) was used to estimate a low-dimensional QSPR model with 12 latent variables. The predictive performance of this model, within the domain of application, was estimated at n=560, q2Ext=0.968 and s=0.028 (log transformed values). The QSPR model was subsequently applied to a database of 100,000+ structures, after a similar 3D optimisation and descriptor generation. Reliable predictions can be reported for compounds within the previously defined applicability domain.

Comparison of global and mode of action-based models for aquatic toxicity

“…Global approaches to modelling aquatic toxicity include k nearest neighbours (kNN) [8,9] or nearest neighbours [10], support vector machines (SVM) [11], multilinear regression (MLR) [10,[12][13][14], MLR using only structurally similar chemicals from the training set [15], group contribution methods [16,17], partial least squares [18,19], artificial neural networks (ANNs) [12,20], associative neural networks (ASNN) [21] and hierarchical clustering (HC) [10,22]. The advantage of global methods is that machine learning allows the development of model(s), which do not require the determination of chemical class or MOA.…”

Global and Local PLS Regression Models to Predict Vapor Pressure

“…It should also be noted that there is a substantial variation in the vapor pressure at environmentally relevant conditions, both seasonal and regional variations (between tropical -subtropical, temperate, and polar regions). Using this model together with our previous model for the enthalpy of vaporization makes it possible to estimate the vapor pressure also at these conditions [51]. …”

Comparison of theoretical and experimental models for characterizing solvent properties using reversed phase liquid chromatography

“…The retention factors for three columns can be analyzed simultaneously, the systematic deviation can be diagnostic by model error e. PLSR can tolerate collinear and missing data, and predictions were still stable for a small number of model dimensions [39]. The QSRR model indicates that a direct chemical interpretation of the latent variables is possible [40]. The dominant factors underlying the PLSR model are related to the molecular volume, polarity, and polarizability.…”